Ecological Economics 180 (2021) 106883

Contents lists available at ScienceDirect

Ecological Economics

journal homepage: www.elsevier.com/locate/ecolecon

Landscape-level feedbacks in the demand for transgenic pesticidal corn in T the Philippines ⁎ Zachary S. Browna,b, , Lawson Connorc, Roderick M. Rejesusa, Jose M. Yorobe Jrd a Department of Agricultural & Resource Economics, North Carolina State University, United States of America b Genetic Engineering & Society Center, North Carolina State University, United States of America c Department of Agricultural Economics and Agribusiness, Louisiana State University, United States of America d University of the Philippines at Los Baños, Philippines

ARTICLE INFO ABSTRACT

JEL codes: We introduce a novel econometric approach to estimate economic pest control feedbacks within agroecological C33 systems, using discrete choice endogenous sorting models. We apply this approach to deployment of transgenic C35 Bt maize in the Philippines. We show with basic theory how areawide pest suppression from largescale Bt maize C36 deployment attenuates farmers' demand for this technology. Econometric results support this hypothesis and D24 imply long-run demand for the Bt trait is price-inelastic, contrasting with price-elastic demand estimated from a Q12 model without feedback. Investigating whether this feedback truly derives from areawide pest suppression, we Q57 analyze farmers' pest infestation expectations and find expected damages are significantly reduced by higher Keywords: areawide Bt deployment. We discuss implications of these findings and other potential applications of the Agroecology econometric approach to study coupled biological and economic systems. Bioeconomic feedbacks Areawide pest suppression Crop choice Discrete choice models Endogenous sorting

1. Introduction genetic composition (e.g. pesticide resistance) of pest populations (Gould et al. 2018). Largescale agricultural production involves ecological feedbacks Theoretical and simulation-based models in biology and economics between farmer decisions, their effects on biological systems, and suggest that externalities resulting from such agroecological feedbacks economic reactions to these effects. The control of crop pests provides a would impact farmers' incentives and decisions to undertake pest con­ particularly salient example. Credible estimates put global crop losses trol efforts (Milne et al. 2015; Epanchin-Niell and Wilen 2015). How­ due to pests at roughly a quarter of potential yields (Oerke 2006; ever, no empirical research has yet investigated whether this is the case. Culliney 2014), with pest damages expected to increase with climate This is at least partly due to the difficulty of such an exercise. The change (Deutsch et al. 2018). Pest control choices strongly influence econometric measurement of such feedback loops would require lar­ economic returns from crop production (Fernandez-Cornejo et al. 2014; gescale, landscape-level experiments or – barring these – some strategy Chambers et al. 2010). A growing body of entomological and weed to address endogeneity. Moreover, directly observing the interactions science research shows how these choices can impose externalities on between farmers and pest populations requires biological and economic neighboring farms, by altering landscape-level pest pressure (Hutchison data with enough detail and variation at both the farm and landscape et al., 2010; Wan et al. 2012; Dively et al. 2018),1 background presence scales. and movement of chemical pesticides (Werle et al. 2018), and the In this paper we introduce a novel approach to econometrically

⁎ Corresponding author at: 4310 Nelson Hall, 2801 Founders Drive, Raleigh, NC 27607, United States of America. E-mail address: [email protected] (Z.S. Brown). 1 Certain pest control choices (like use of transgenic Bt crops) not only alter landscape-level pest pressure for the primary (e.g., target) pest, but these choices can also alter landscape-level pest pressure for secondary (e.g., non-target) pests. For example, one would generally expect that landscape-level pest pressure for the primary pest would decrease with Bt corn and then landscape-level pest pressure for the secondary pest (not susceptible to the Bt toxin) may increase (Kuosmanen et al. 2006; Catarino et al. 2015). https://doi.org/10.1016/j.ecolecon.2020.106883 Received 31 December 2019; Received in revised form 6 October 2020; Accepted 7 October 2020 0921-8009/ © 2020 Elsevier B.V. All rights reserved. Z.S. Brown, et al. Ecological Economics 180 (2021) 106883 measure the effect of bioeconomic pest control feedbacks on agents' systems is the Asian corn borer (Ostrinia furnicalis, or ACB). Bt varieties choices, by using endogenous sorting models developed in the en­ are also highly effective at suppressing ACB (Afidchao et al. 2013). vironmental, resource and urban economics literatures (Bayer and We first estimate an endogenous sorting model for Philippine maize Timmins 2007; Timmins and Murdock 2007; Hicks et al. 2012; farmers' decisions to deploy Bt versus non-Bt varieties, using a two- Bernasco et al. 2017).2 These tools have been specifically developed to stage fixed effects conditional logit and instrumental variables regres­ measure how the choices of individual economic agents depend on the sion method originally developed by Bayer and Timmins (2007). The market-level, aggregate demand shares for discrete alternatives, and in results from our estimation confirm the prediction of our conceptual turn how these endogenous choices feed back into determining ag­ model: Greater areawide adoption of Bt varieties appears to attenuate gregate demand. For example, demand for recreation areas is likely to individual demand for them. As noted above, in qualitive terms, the attenuate with increased congestion (Timmins and Murdock 2007). The endogeneity appears to generate price-inelastic long-run demand for key analogue we use here is the hypothesis, based on previous agroe­ the Bt trait. To corroborate the bioeconomic explanation for this result, cological research, that demand for a specific pest control technology – we then econometrically analyze survey data on farmers' expectations genetically modified Bacillus thuringiensis (Bt) maize – should attenuate regarding ACB infestations, to see whether higher areawide Bt de­ due to landscape-level pest pressure reductions induced by the tech­ ployment leads farmers to expect fewer infestations. We find robust nology. evidence for this effect. The general identification strategy used in endogenous sorting The rest of the paper proceeds as follows: We first present a con­ models is to observe variation in the effective choice set available ceptual model of pest suppression spillovers and feedbacks in the de­ within each market. Such variation can be used to build an instrument mand for Bt maize, followed by a detailed description of our econo­ that predicts variation in aggregate demand shares, but is exogenous metric approach. We then provide a description of the empirical context with respect to unobserved market-level factors that affect individual and the dataset used in the study, before discussing some econometric preferences and thereby aggregate demand. For example, recreation considerations (vis-a-vis our data). Estimation results are then pre­ site openings or closures – or more generally changes in the relative sented and interpreted. We then discuss implications and limitations of costs of access or nonmonetary attributes – over time produces varia­ our analysis. tion in the effective recreation choice set available at a given interval of time (Timmins and Murdock 2007). To the extent this variation is 2. Conceptual and Econometric Models exogenously determined (i.e. not itself a product of congestion), it can be used to infer how congestion attenuates demand. We first present a conceptual model of how we can expect area-level To apply this approach in measuring the bioeconomic pest control adoption of a pesticidal crop to determine pest densities (the bioeco­ feedbacks, we first present a simple conceptual model showing how an nomic spillover), and in turn determine individual grower choices analogous feedback should manifest in the case of Bt maize, given about whether to adopt GM varieties. While our model makes a number reasonable – and testable – assumptions about agents' pest damage of simplifications, in the following section describing the study area, we expectations. In this model, greater areawide adoption of Bt maize re­ argue that it characterizes well, the typical insect pest control en­ duces landscape-level pest pressure, thereby reducing the incentive for vironment facing a Philippine maize farmer over the time period in our individual farmers in the area to deploy the control measure. One data. We then translate this conceptual model into an econometric practical consequence of recognizing this bioeconomic feedback is that approach, and describe the estimation procedure. the long-run, equilibrating price elasticity of demand for the pest con­ trol measure should therefore be less than what would be erroneously 2.1. Conceptual Model implied by ignoring it. Our empirical estimates in fact suggest demand for the Bt trait switches from being price-elastic to inelastic when ac­ Consider a farmer facing the ex ante binary choice of whether to counting for this endogeneity. plant one of two varieties of a crop: a conventional variety fully sus­ Motivated by this theory, we then specify and estimate an en­ ceptible to pest damage or a pesticidal variety that protects the plant dogenous sorting model of pest control decisions, using data on the from damage and also kills the pest (as is the case with Bt maize). To fix deployment of genetically modified, Bt maize in the Philippines. Bt ideas with respect to our application to Bt maize, we refer to the con­ crops express insecticidal toxins (which, in nature, are produced by Bt ventional variety as the hybrid (H) and the pesticidal variety as the Bt bacteria) that are highly effective at killing and preventing damage variety. In the model, farmers do not observe pest densities in the from economically important pests. A largescale analysis of Bt maize coming season, but have expectations about future pest pressure (e.g. deployment in the Midwestern U.S., by Hutchison et al. (2010), showed based on previous years and on forecasts of environmental conditions). that European corn borer (Ostrinia nubilalis, or ECB) – once one of the For simplicity, our conceptual model focuses only on uncertainty most historically significant pests of maize in the region – was effec­ with respect to pest densities in the upcoming season. Let πH(d) be the tively eliminated by the widespread deployment of Bt varieties. In ex post profit (crop revenues less costs of seed and other inputs) from particular, Hutchison et al. (2010) showed that widespread use of these the non-Bt hybrid given a pest density of d, and πBt the ex post profit varieties decreased abundance of the pest on both Bt and non-Bt maize from adopting the pesticidal variety, apart from the price premium for over the landscapes studied. Based on this observation, they estimated the Bt variety. Assume that πH′ < 0, i.e. that ex post profit from the the majority of the net benefits from eliminating this pest were accrued hybrid variety is decreasing in pest density, and that the pesticidal crop by non-users of Bt varieties, who avoided paying the additional costs for is fully protected against pest damage so that πBt is independent of pest the proprietary seeds. Subsequent entomological research in a variety density. Also, suppose that given an areawide Bt adoption level of of agricultural systems has added to the evidence for areawide pest C ∈ [0,1] the ex ante cumulative distribution function (CDF) for d is F suppression by Bt crops (Wan et al. 2012; Dively et al. 2018). In the (d|C), which defines farmer expectations about pest densities d in the context of our study, a key related pest in Philippine maize farming upcoming season, conditional on areawide adoption C of the Bt variety. Finally, let w denote the price premium for the Bt variety. Then ex ante expected profits for the hybrid and Bt varieties are: 2 Throughout this paper, we use the term bioeconomic as an abbreviation for H (C )d [ H (d ) | C ] (1) ‘coupled biological and economic’ (system). This mirrors the term's meaning in the context of fisheries (Clark 2010), although we do not analyze explicit bio­ Bt d[ Bt w | C ] = Bt w logical data in this paper (beyond farmers' expectations about pest infestation) in the way that harvest data are used for bioeconomic analyses of fisheries. where the operator d[·|C ] emphasizes that we are focusing on

2 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883 uncertainty with regard to pest densities conditional on an assumed level of areawide Bt adoption. The farmer will therefore adopt the Bt variety if πBt − ΠH(C) > w and will plant the conventional variety if πBt − ΠH(C) < w. That is, the farmer will base the decision on the ex ante expected profit differential ΔΠ(C) ≔ πBt − ΠH(C), relative to the price premium w of the Bt variety. A generic way to model a pest suppression effect of areawide adoption in the above framework is to assume that F(d|C) > F(d|C′) for all C > C′, i.e. the CDF conditional on C′ first-order stochastically dominates any CDF conditional on a higher C. Under this assumption, and because πH(d) is strictly decreasing, then ΠH′ > 0 (a basic im­ plication of first-order stochastic dominance). Consequently, the­ ex pected profit gain from the Bt variety relative to the hybrid variety is decreasing in areawide adoption, i.e. ΔΠ ′ < 0. This provides an intuitive model of a negative, pest suppression feedback from pesticidal crop adoption. As areawide use of Bt increases, farmers expect decreased pest pressure on their own farms.3 Equili­ brium properties mirror those found with respect to congestion ex­ ∗ ternalities (Bayer and Timmins 2005): If any solution C to the equation ∗ ΔΠ(C ) = w exists on the interval [0,1], then it is the unique equili­ Fig. 1. Illustration of pest suppression feedbacks in Bt seed choices. brium of the model, the point at which the marginal farmer is in­ different between adopting Bt or the conventional variety. This equi­ With the pest suppression feedback and this random component to librium is stable in the sense that there is an individual incentive to ∗ profits, an equilibrium areawide Bt adoption level C now satisfies DBt ∗ ∗ adopt Bt if areawide adoption is below equilibrium, and disincentive to (w,C ) = C . Applying the Implicit Function Theorem to this equation adopt if areawide adoption is above equilibrium. That is, ΔΠ(C) > w ∗ ∗ ∗ to derive dC /dw when ΔΠ′ < 0 (i.e. areawide pest suppression) and for all C < C and ΔΠ(C) > w for all C > C . Fig. 1 illustrates such an given f(·) > 0, we obtain the relation: equilibrium. If no solution to this equation exists, then ΔΠ(C) exceeds w D f[ w()] C across the whole unit interval, in which case full adoption of Bt is the Bt = f[ w()] C < equilibrium, or w exceeds ΔΠ(C) across the whole unit interval, in w CC= 1f [ w()] C which case the unique equilibrium is no adoption of the Bt variety. () DBt Alternatively, a pest control method could feasibly result in repel­ dC = w =( < 0) ling – rather than suppressing – pests from areas where the method is D Bt 1 dw adopted to areas where the method was not yet adopted. In this case C CC= (2) CDFs of pest pressure for non-adopters conditional on high area-level adoption could stochastically dominate those with lower adoption, ul­ This relation shows the price elasticity of demand accounting for timately flipping the polarity of the modeled feedback from negative to pest suppression feedbacks is lower in absolute magnitude than the dC w < DBt w positive. This would be analogous to an agglomeration externality price elasticity ignoring it, i.e. dw C w C . In economic terms, a (Bayer and Timmins 2005). In our context, we hypothesize that a pest price change in the Bt variety is met by a countervailing change in pest suppression feedback is the most relevant for study, rather than re­ pressure which attenuates the demand effects of the price change. pulsion. For Lepidopteran pests like ECB and ACB, Bt crops have a clear suppressing as opposed to repelling effect because they act on the ca­ 2.2. Econometric Approach terpillar larvae, which have limited ability to avoid exposure to the Bt toxins (Hutchison et al., 2010). To empirically evaluate the presence of endogenous feedbacks from Such an areawide pest suppression feedback will also in general Bt seed use, we use an IV method developed by Bayer and Timmins reduce the absolute price elasticity of demand for the Bt variety. To see (2007) to estimate discrete choice econometric random utility model this in a discrete choice model of seed choice, we must extend the above (RUM) with endogenous sorting. We apply this method to model model to allow for heterogeneity or some random component to farmer farmers' crop variety choices. Following Ciliberto et al. (2019), who profits. For exposition assume that the profit differentiali ΔΠ (C) for also estimate a discrete choice model of farmers' seed variety choices ≔ ϵ ϵ farmer i is ΔΠi(C) ΔΠ(C) + i, where i is a random component of the (without endogenous sorting), ex ante utility can be interpreted as im­ ϵ farmer returns to Bt-versus-hybrid varieties, with pdf f( i). Then all plicitly containing the expected profit from selecting seed varietyj , but ϵ ϵ farmers with i such that ΔΠ(C) + i > w will adopt the Bt variety, and may also be related to other factors directly affecting utility, such as D(,)( w C f) d the resulting demand for the Bt variety is Bt w i i. farmer preferences specifically regarding genetically modified crops (Useche et al. 2009; Birol et al. 2008).4

3 However, as one reviewer has noted, areawide pest suppression benefits of Bt corn not only depends on the relative share of Bt corn on total corn pro­ 4 Other factors that may influence utility from adoption of Bt seed variety are duction area, but also on the share of total corn production relative to the total existing policies regarding refuge requirements and coexistence regulations cropland area. In the Philippines in 2018 (and in 2007 when the first survey we (Skevas et al. 2010; Groeneveld et al. 2013). For the study period in this paper, analyze was conducted), corn was the second most important crop after rice there was a 20% refuge requirement in place in the Philippines (Rodriguez (see section 3) (FAOSTAT 2020). Moreover, between 70%–100% of crop acres 2014), though there is no reliable information with regards to enforcement and is planted to corn in the lowland and 35%–100% in the uplands (Gerpacio compliance to this directive in the study period. In countries where they are 2004). Our study areas also have traditionally agricultural landscapes where present, Bt refuge mandates or recommendations are often poorly enforced, the proportion of agricultural land to total area ranges from 44% to 60% particularly in areas with relatively small, heterogeneous farms and larger (Philippine Statistics Authority, 2014). Hence, the relative share of corn in the numbers of farmers (Carrière et al. 2020). In addition, note that there are no country and the proportion of agricultural land to total land area supports the coexistence policies currently in place in the Philippines (nor were there at the existence of potential area-wide pest suppression and indicates that this con­ time of the study). Thus, these policy-related factors likely do not play a major ceptual model applies to our empirical setting. role in affecting utility derived from Bt seed adoption. One reviewer also noted

3 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883

We specify the ex ante utility Ujih to the farmer of crop variety j for different varieties that make them more or less suited for a given area, grower i in area h by partitioning farmer and area-level utility com­ (ii) unobserved expected profit and utility effects of other inputs (e.g. ponents: fertilizer, pesticides, or labor) that are altered by adoption of Bt-versus- alternative maize varieties (Afidchao et al. 2013), and (iii) local eco­ Ujih= jh+ x ji+ jih (3) logical conditions favoring pests that may increase relative demand for pesticidal varieties. where δjh is the area-level effect of variety j on utility, xji is a vector of farmer-level covariates varying across varieties, β is an associated To specifically account for endogenous sorting in the econometric analysis, Bayer and Timmins (2007) propose an IV for the market shares vector of regression coefficients, andϵ jih is a random utility component. The area-level effect is decomposed as: σjh. In their originally intended application to urban economics, j in­ dexes geographic location (rather than seed variety) and h indexes = p +Bt C + jh j jh j h jh (4) market. In the context of geographic sorting models, they propose as an ≔ IV a function f(Xj,X−j) of “the attributes [Xjh (xjih)i=1, …, nh] of lo­ where j is a variety-specific constant,p jh is the price of variety j in area cation j and the exogenous attributes [X−j ≔ (x−jih)i=1, …, n ] of other η h h with associated marginal utility , Ch is the fraction of growers in area locations [−j]” (pp. 361, emphasis added). In practice, this function is h employing varieties with the Bt trait, Btj is a dummy variable in­ the conditional logit predicted probability, dicating whether variety j possesses the Bt trait, and ξjh is an area-level 1 exp xjih f( Xjh, X jh) nh i h . Bayer and Timmins demonstrate residual. We aim to estimate the utility parameters α, β, η and, . k h exp xkih The parameter we focus on is the spillover effect, α, of areawide Bt the validity and performance of this estimator using Monte Carlo ana­ lysis, and a number of subsequent studies have applied this method to use, Ch. Note that areawide Bt use Ch is interacted with Btj in eq. (4) for both theoretical and mechanical reasons. The simple theory in section domains ranging from congestion in fishing sites (Timmins and 2.1 implies that greater areawide use of Bt decreases the relative utility Murdock 2007; Hicks et al. 2012), nonmarginal valuation of climate amenities (Timmins 2007), to urban crime (Bernasco et al. 2017). (∂Ujih/∂Ch < 0) only for those j with the Bt trait (Btj = 1), yielding the Bayer and Timmins (2007, p. 365) explain that identification using hypothesis that α < 0. Mechanically, Ch cannot enter alone as a cov­ ariate in (4) because this variable does not vary over alternatives j, and their instrument is provided by exogenous variation in “effective choice it is only differences between alternatives that identify preferences in a sets” – that is, not only variation between areas in the available alter­ natives (seed varieties in our application), but also exogenous variation RUM (Train 2009). Interacting Ch with Btj produces necessary variation over j. in the relative utility those alternatives convey. In our application this means that exogenous, between-area differences in the utility conveyed Assuming the farm-level random utility component ϵjih in (3) is iid extreme value, we obtain the fixed effects conditional logit model for by each seed varietiy can generate variation in the effective choice set. For example, more remote farmers might find it more costly to access the probability Pjih of grower i selecting variety j in area h: agricultural inputs such a pesticides and thus find the built-in pest exp{xji + jh } control services of Bt seed varieties relatively more attractive. In this Pjih(,) h = exp{x + } k h ki kh (5) example, we would then expect relatively greater exogenous utility from the Bt trait in areas characterized by relatively remote farms. In­ where “k ∈ h” is short-hand to indicate the denominator in (5) sums teracting an indicator for farm remoteness with the variety-specific Bt over all the varieties k available in area h (in our application, areas are trait would allow for such observable preference heterogeneity to village×year). generate effective choice set variation. Similar conceptual examples The standard approach to estimating this model is via a two-stage could be constructed to illustrate how observable heterogeneity in procedure. In the first stage, estimates and h() jh j=1, … , J are ob­ preferences for the herbicide tolerance (HT) trait (e.g. via farm terrain) tained from maximum-likelihood (ML) estimation combined with a could also generate effective choice set variation. contraction mapping algorithm from Berry et al. (1995). In the second To implement this identification strategy, we include an array of stage, the estimated jh serve as dependent variables in a linear re­ exogenous farm(er) characteristics interacted with variety-specific gression on observable variety-specific factors varying at the area level, dummy variables in the discrete choice RUM (since discrete-choice using the decomposition in (4) and treating the unobserved area-level RUMS require such characteristics to interact with alternative-specific component ξjh as a regression error. If the explanatory variables in (4) variables). Our IV for areawide Bt deployment therefore takes the fol­ ξ are orthogonal to jh, then this second-stage can be estimated con­ lowing form in our application: sistently with OLS or other standard panel data methods (Murdock Btj exp{ xjih pjh } 2006). C IV h j h nh exp{ x p } However, with endogenous sorting, area-level explanatory variables i h k h kih kh (6) in the second stage include the area-level adoption of Bt. This creates an obvious endogeneity problem, since correlation in areawide un­ where is an initial guess of the price coefficient obtained by regres­ observables likely implies ( jh | Ch) 0. As Timmins and Murdock sing on pjh without areawide feedbacks (Timmins and Murdock (2007) point out, naïve OLS of (4) tends to bias estimates of α upwards, 2007; Hicks et al. 2012; Bernasco et al. 2017). because of unobserved area-level factors giving rise to correlated One potential limitation of the above methodology is the condi­ choices. Examples of such unobserved areawide correlation in the tional logit model's assumption of the ‘independence of irrelevant al­ present context include (i) unobserved agronomic characteristics of ternatives’ (IIA) (McFadden 1978), which would imply that introduc­ tion of the stacked HT-Bt variety would draw farmer demand in equal proportions away from the other available varieties (the single-trait Bt (footnote continued) and non-GM hybrid maize). Although our inclusion of area-level fixed the potential role of option values in farmers' seed decisions, e.g., the “sunk” effects in (3) relaxes the IIA assumption at the area level, as a robust­ nature of Bt seed purchases vis-à-vis the value of waiting for information about ness check we also estimate a mixed logit model with area-level fixed the likely severity of pest infestation in the upcoming season. While we do not effects (see Supplementary Material). We do not reject the more par­ address this aspect explicitly in our analysis (i.e. we cannot estimate this option value), it does not change the main implications of our conceptual model in simonious conditional logit model. Section 2 (by including this option value as a discount on the expected relative As an additional investigation of bioeconomic feedbacks, we in­ benefits of adopting Bt seed), and can be viewed as being subsumed within our vestigate whether results from our RUM are consistent with other in­ econometric model. See Mbah et al. (2010) for analysis of this dimension of the dicators of areawide pest suppression. Areawide pest suppression from decision problem in bioinvasions. Bt implies that higher areawide deployment of Bt decreases expected

4 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883 pest pressure. While lacking pest monitoring and entomological data, The data used in this study come from the International Food Policy our surveys elicited a binary indicator bih of whether farmer i in village- Research Institute (IFPRI) maize surveys for crop years 2007/2008 and year h expects ACB infestation. To test for areawide ACB suppression, 2010/2011 in the Philippines. The data represent a panel where 278 of we therefore estimate the following probit regression: the farmers in the 2007 cycle were retained into 2010. Data collected in the survey included information on maize farming systems and en­ [bih | C h , c ih , X ih ] =( +Ch +cih + Xih) 0 C c X (7) vironment, inputs and outputs, costs and revenues, marketing en­ vironment, and other factors related to Bt maize cultivation were col­ where cih is an indicator of whether farmer i plants a Bt variety (single lected (i.e., subjective perceptions about the technology). Actual data trait or stacked), Xih is a vector of control variables, Φ(·) is the standard normal CDF and the β’s are regression coefficients to be estimated. To collection was implemented through face-to-face interviews using pre- tested questionnaires. allow cih to be endogenous, we estimate (7) using a full-information ML (FIML) bivariate probit regression (see Supplementary Material for The survey was confined to the provinces of Isabela and South details). Cotabato, both major maize-producing provinces with a high, sustained level of Bt crop deployment. The non-Bt farmers in our data are strictly hybrid maize users, and there were no observations in the data of 3. Study Context and Data farmers using traditional, open-pollinated varieties. This uniformity in the non-Bt group allows for a useful baseline to compare the demand for We apply the above econometric framework using data from sur­ Bt maize compared non-Bt farmers (i.e. hybrid maize users only). veys of Filipino maize growers. Maize is the second most important crop Seventeen top maize-producing villages (‘barangays’) were selected for in the Philippines after rice, with approximately one-third of Filipino surveying from these two provinces. Survey sampling proceeded by farmers (~1.8 million) depending on maize as their major source of obtaining lists of farms from each village, and randomly selecting a livelihood. Yellow maize, which accounts for about 60% of total maize fixed proportion of farms for surveying. production (white maize accounts for the rest), is the type considered in A total of 468 farmers were interviewed in the 2007/2008 round this study. Maize growing in the Philippines is typically rain-fed in and 278 of those farmers were also interviewed in the 2010/2011 lowland, upland, and rolling-to-hilly agro-ecological zones of the round of data collection. After dropping farmers with missing and in­ country. There are two cropping seasons per year: wet season cropping consistent information, a total of 683 total observations across both (usually from March/April to August) and dry season cropping (from survey years. For the purposes of this analysis, we furthermore exclude November to February). Most maize farmers in the Philippines are villages with fewer than eight growers, due to the difficulties of esti­ small, semi-subsistence farmers with an average farm size ranging from mating the δjh’s in (5) with such small area-level sample sizes. In re­ less than a hectare to about 4 ha (Mendoza and Rosegrant 1995; tained villages, we also restrict econometric analysis presented here to Gerpacio et al., 2004). the balanced panel of 261 growers present in both the 2007 and 2011 The most destructive pest in the major maize producing regions of surveys. We focus on the balanced panel because of additional in­ the Philippines is ACB (Morallo-Rejesus and Punzalan, 2002; Gerpacio formation that was collected in the 2011 survey and which we use in et al., 2004; Afidchao et al. 2013). Like its European relative, ACB the analysis here, such as the distance of the farm to the nearest road.6 larvae damage all parts of the maize plant in feeding, before meta­ Table 1 summarizes the adoption shares for the different seed types morphosing into moths, which can disperse widely (Nafus and by village (corresponding to the σjh in section 2.2). From this we can Schreiner 1991; Shirai 1998). Historically, ACB infestation has occurred quickly see a number of patterns. First, there is significant hetero­ yearly, with pest pressure being roughly constant or increasing over geneity in GM crop adoption between villages and years. Second, be­ time. Farmers report that yield losses from this pest range from 20% to tween 2007 and 2011 there was a significant shift to GM varieties, 80%. According to Gerpacio et al. (2004), although ACB is a major pest specifically to the stacked trait variety. In particular 100% ofthe in the country, insecticide application has been moderate compared to sampled farmers in five of the 11 villages in 2011 chose the stacked- other countries in Asia. Insecticide application for control of ACB is trait variety; this evidently high demand both poses complications and difficult to effectively implement, often not economical, can interfere offers some identifying variation for our proposed econometric­ ap with beneficial parasitoids, and can have adverse health and environ­ proach. mental effects Nafus( and Schreiner 1991; PestNet 2017). In the Phi­ To estimate the choice models used in this study, we require subsets lippines, insecticide applications for ACB have been uncommon (Nafus of variables that differ over area, individual and variety. Identification and Schreiner 1991; Yorobe and Quicoy 2006). requirements for these variables are that they should be exogenous to Given ACB's dominance as the major insect pest for maize in the both individual choices and area-level adoption of GM varieties. Table 2 country, the agricultural sector was naturally interested in Bt maize summarizes the grower-level variables used in this analysis. At the in­ varieties as a means of control. As with ECB, Bt maize is highly effective dividual level, we include individual growers' distances to the nearest 5 at suppressing ACB larvae (Afidchao et al. 2013). In December 2002, seed supply source and nearest road in the first-stage estimation, fol­ after extensive field trials, the Philippine Department of Agriculture lowing Sanglestsawai et al.'s (2014) study of the yield effects of Bt (DA) provided regulations for the commercial use of GM crops, in­ adoption in the Philippines using the 2007 survey data. We also include cluding Bt maize (specifically Monsanto's YieldgardTM 818 and 838). In a measure of farmer experience – the number of years farming maize (as the first year of its commercial availability 2002, Bt maize was grown in of 2007) – and basic indicators of the farm terrain. only 1% of the total area planted with maize – on about 230,000 ha. In While we do not observe pest densities, we do use in this analysis 2008, about 12.8% of maize planted was Bt, and in 2009 this increased survey data on farmer expectations about the future ACB infestations. to 19%, or about 500,000 ha. Apart from Monsanto, Pioneer Hi-Bred Table 2 also shows the mean and standard deviations for responses to (since 2003) and Syngenta (since 2005) sell Bt maize seeds in the the survey question: Philippines. In using this variety [of seed selected by the farmer], do you expect corn borer infestation? (Yes/No).

5 We employ responses to this question as the pest infestation Other Bt varieties, expressing proteins for controlling rootworms, were not available in the Philippines over the time period analyzed, because these are not considered major pests of corn there. The only other common non-ACB insect pests of corn in the Philippines are the armyworm and cutworm 6 We have also replicated the analysis with the unbalanced panel of farmers, (Gerpacio et al. 2004; Afidchao et al. 2013), which the available Bt varieties do excluding the variables only collected in 2011. The main results of the paper are protect against. robust; results are available on request to the authors.

5 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883

Table 1 Table 3 Maize variety adoption shares and number of surveyed growers by village. Variety-specific, area-level seed prices (Philippine pesos, PHP).

2007 2011 2007 2011

Province Village / Hybrid Bt N Hybrid Bt Stacked N Variety Mean Std. Dev. Mean Std. Dev. Barangay Conventional hybrid 185 32 274 36 Mindanao Olympog 71% 29% 38 14% 18% 68% 28 Bt single-trait 300 44 386 48 Sinawal 79% 21% 52 65% 27% 8% 26 Bt/HT stacked-trait n/a n/a 451 42 Tampakan 73% 27% 70 27% 9% 64% 22 Isabela Andarayan 30% 70% 10 0% 0% 100% 8 Table 3 notes: These data are obtained from an OLS regression of seed prices Bugallon 46% 53% 28 0% 17% 83% 18 paid by growers on village-level fixed effects interacted with variety-specific San Pablo 50% 50% 20 0% 0% 100% 14 dummy variables and an independent time trend. Prices for stacked trait in Villa Luna 26% 74% 35 0% 20% 80% 20 2007 are not applicable (n/a) because this variety was not available in that Cabaseria 5 29% 71% 92 0% 0% 100% 60 year. Dappat 45% 55% 33 0% 0% 100% 22 San Fernando 28% 72% 36 3% 0% 97% 34 San Manuel 7% 93% 14 0% 0% 100% 12 summarizes these computed variety-specific prices. The price premium TOTAL 207 221 428 28 21 215 264 for Bt single-trait in 2007 is 62% that of the mean conventional hybrid 48% 52% 11% 8% 81% price, declining to 41% in 2011. The premium for the stacked variety is 65% of the mean hybrid seed price in 2011. The price of the hybrid variety increased by an average of 48% between 2007 and 2011. These Table 2 relative premiums for single-trait Bt and stacked trait varieties are Grower-level characteristics used in the choice models. roughly in line with those reported by Ciliberto et al., 2019, Table 3) for Mean Standard Village-level Village-level the U.S. corn seed market. The price premium differential is largely deviation std. dev.1 variation (%)2 expected, e.g. given the large regulatory costs involved in getting Years maize farming 22 11 4 36% varieties with genetically-modified traits approved by country govern­ Distance to roads (km) 0.5 1.1 0.3 31% ments (Smart et al. 2017). Given that the costs of Bt seed were esti­ Distance to seed source 6.2 10.2 3.5 34% mated at that time to amount to around a fifth of Philippine maize (km) growers' input costs (Afidchao et al. 2013), these premiums can be Terrain Flat 66% 48% 29% 61% expected to comprise an important consideration in their seed choices. Rolling 21% 40% 17% 42% Hilly or mountainous 14% 35% 14% 40% 4. Econometric Estimation and Specification Expect corn borer 45% 50% 38% 76% infestation The most significant complication in implementing our econometric Table 2 notes: 1. Standard deviation in village×year-level means, 2. Defined as approach with these data is the presence of some 0% and 100% village- the standard deviation of area (village×year) means divided by the total level adoption shares for 2011, which causes estimated jh’s in these standard deviation. cases to converge to negative or positive infinity. To deal with this issue, in addition to an IV quantile regression (IVQR) used by Timmins indicator in regression eq. (7). As noted above, this indicator is clearly and Murdock (2007) for this same purpose, we estimate an IV Tobit in endogenous with seed choice. Table 2 shows that perceived pest in­ the second stage regression. Whereas the IVQR treats the 0% or 100% festation risk exhibits proportionally greater between-village than village-level adoption as outlier observations, the IV Tobit more ap­ within-village variation, more than any of the other variables in this propriately treats them as censored observations. The censoring in this table. This suggests a high within-village correlation of this variable and application results from being unable to observe finite ’s in cases an important areawide component to perceived pest infestation risks. where the finite survey sample produces no variation in choices within Lastly, to obtain variety-specific prices which vary over villages and these areas. The details of how we implement this approach and eval­ years for the RUM in (3) and (4), we follow Ciliberto et al. (2019) in uate its validity are described in the Supplementary Material. their discrete choice model of seed choice and compute the average Another econometric issue for estimating the RUM specified in (3) price for each maize variety in the dataset for each area. Like in Cili­ and (4) is the potential for seed price endogeneity (Ciliberto et al. berto et al., farm-level seed prices are only observed for purchased 2019). For example, within-area market power might permit seed varieties, and so we regress the survey-elicited price φivt that farmer i producers to increase prices to capture rent, in a way that varies sys­ paid for their seed in village v, year t on village fixed effects interacted tematically across local geographies or between years (e.g. due to the with the seed type planted by the farmer and a year dummy. 2007/2008 world food price crisis, as a reviewer noted). This concern is somewhat alleviated because the logic of RUMs implies only between- ivt =()jv + jtc ijt + ivt variety differences (pjh − pkh) in seed prices affect choices in these j (8) models. Any endogenous markups to price would have to differ by variety and by area to contaminate our model with price endogeneity. where cijt indicates which variety j farm i purchased in year t, the θ’s are For example, if an omitted factor was driving endogenous variation in regression coefficients to be estimated and υivt is the residual. After estimating (8) via OLS, we use predictions from this regression to obtain GM seed prices but this variation was constant across varieties, then it area-level prices, where an “area” is defined here and throughout as a would have no effect on the identification of a discrete choice RUM. 7 year-village combination h = (v,t), so that pjh =jv + jt. Table 3 Alternatively, if there was endogenous variation in GM seed price premiums that was constant across areas, it would be picked up by the variety-specific constants ( j in eq. 4), which we make no attempt to 7 Note that in a departure from Ciliberto et al. (2019), who simply calculate economically interpret in this manuscript (e.g. for an overall economic the market-level variety specific averages prices for each variety, we specify village θjv and time θjt effects as additively separable. This is specifically be­ cause of the 5 villages in 2011 for which we only observe farmers purchasing (footnote continued) the stacked variety (Table 1), therefore making estimation of a fully model with prices for the Bt single trait and hybrid varieties even in village-years where no village×year fixed effects infeasible. That is, we require imputations for the one was observed purchasing these varieties.

6 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883 valuation of GM maize, which is a central aim of Ciliberto et al.'s for use of the GM seed varieties. As in Sanglestsawai et al. (2014), analysis). Nevertheless, heterogeneous GM-specific premiums remain distance to nearest seed source appears in these columns to have a possible. Ciliberto et al. (2019) address price endogeneity in the U.S. counterintuitive effect on use of the single-trait Bt variety, with farms seed market by instrumenting with the total number of competing farther from their nearest seed supplier evidently more likely to pur­ products in local markets. We cannot execute this method in our data, chase the single-trait Bt seed. However, the counterintuitive effect of because in any given year the number of competing products is constant seed supplier distance on single-trait Bt use becomes insignificant when across villages. area-level fixed effects are included (second and fourth columns of We therefore adopt a two-part approach to address the potential for Table 4). price endogeneity vis-à-vis our primary empirical aims. We use our The farm's distance to the nearest road also appears to significantly knowledge of the Philippine maize seed market and a statistical check explain demand for the single-trait Bt variety (Table 4). The positive for systematic variation in prices. As noted in Section 3 above, the seed sign of the estimated coefficient, in all specifications, implies more re­ brands available in the villages in our data did not change over the mote farms (i.e. those less connected to transportation networks) are course of the dataset. Also as noted above, single trait Bt was uniformly more likely to use the single-trait Bt variety (as well as the stacked trait available in the sampled villages in 2007, and both the single trait and variety, though none of these coefficients are statistically significant). stacked varieties were uniformly available in 2011. This uniformity This patterns appears even stronger when area-level fixed effects are means that we would not expect market power to be differentially ex­ included in the regression (in terms of statistical significance; recall that ercised heterogeneously across villages, which would therefore not the coefficient magnitudes in RUMs cannot be directly compared be­ empirically generate price endogeneity. tween specifications). Multiple explanations could accommodate this We also consider potential market power at the seed retailer level: result: More remote farms may find it more costly to access local labor On this point, our understanding is that there were no significant in­ markets, and be using the Bt and herbicide-tolerant varieties to reduce stitutional barriers to entry for sellers in the area-level retail markets in hired labor inputs in pre-harvest pest control (e.g., detasselling) and the Philippines. However, one possibility is that the size of the area- weeding activities (Felkl, 1988a,b; Gouse et al. 2016; Connor 2017). level seed markets limited entry and affected market power Campbell( Areawide pest suppression may also play a role: More remote farms and Hopenhayn 2005; Melitz and Ottaviano 2008). To investigate this, may also enjoy fewer benefits of areawide pest suppression from we examine whether there is a relationship between pjh and measures of neighbors' use of Bt seed, thereby increasing remote farms' own in­ area-level market size (elicited in our survey). The Supplementary centive to use these varieties, ceteris paribus. Alternatively, more re­ Material provides details on this test, which cannot statistically reject mote farmers may find it more costly to access and reactively apply the null hypothesis of no such relationship. topical insecticides upon pest outbreaks, and thus find the Bt trait re­ latively more desirable as a means of ACB control. This latter hypoth­ esis could also explain why distance to roads has a smaller and insig­ 5. Results nificant effect on utility from the stacked variety, since access to herbicides is important to benefit from the HT trait in this variety. Table 4 shows the first-stage conditional logit models, with and The dummy variable indicators for terrain, despite their coarseness, without area-level fixed effects. For our purposes the important take­ show explanatory power and mostly intuitive relationships with seed away from this table is that the set of farm-level covariates, taken as a choices in Table 4. Farms with “rolling terrain” generally appear more whole, have statistically significant explanatory power over seed choice 2 likely to use the GM varieties over the hybrid. This could be explained (as seen in the significant Waldχ statistics in the table). In the baseline by ecological conditions on such terrain that favor ACB and weeds. conditional logit regressions without area-level fixed effects (first two Afidchao et al. (2013)find that ACB damage on maize in the Philippines columns of Table 4), a farm's distance to the nearest seed source and is positively correlated with distance away from rivers and floodplains. indicators of farm terrain appear to have the most explanatory power

Table 4 First-stage random utility model estimates (conditional logit, see eq. 5).

Conditional logit Fixed-effects conditional logit

Log(seed price) −1.071* [in fixed effect] (0.564) Maize variety: Bt single-trait Stacked Bt single-trait Stacked ×Constant 0.233 2.463*** [in fixed effect] [in fixed effect] (0.385) (0.530) ×Distance to seed source 0.0288*** −0.0490** 0.0149 −0.0714* (0.0105) (0.0217) (0.00967) (0.0409) ×Rolling terrain 0.865** 2.228*** 0.139 1.245* (0.342) (0.733) (0.320) (0.730) ×Hilly/mountainous terrain −0.561 −0.360 −1.005*** −0.875 (0.344) (0.446) (0.352) (0.640) ×Distance to nearest road 0.244 0.102 0.362** 0.111 (0.165) (0.246) (0.153) (0.242) ×Years farming maize 0.00197 0.0289 0.00125 0.0149 (0.0132) (0.0215) (0.00781) (0.0172) Area fixed effects# No Yes Choice occasions 515 515 Farmers 261 261 Deg. Freedom 13 10 Log-likelihood −313.6 −220.9 Wald-χ2 196.66*** 41.11*** Pseudo-R2 0.321 0.0575

Table 4 notes: Robust standard errors clustered at the grower level and in parentheses. Statistical significance: *** p < 0.01, ** p < 0.05, *p < 0.1. # Area-level fixed effects model calculated using contraction mapping algorithm (Berry et al. 1995). Area-level coefficients (price and variety-specific constants) are contained within area-level effects. ‘Within’ pseudo-R2 calculations in fixed-effect models calculated relative to a null conditional logit model with only area-level fixed effects.

7 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883

Table 5 Second-stage random utility model estimates (see eq. 4).

OLS 2SLSa IVQRa,b IV Tobita,c

(1) (2) (3) (4) (5)

Area Bt fraction × Bt variety −7.811** −17.79** −19.98*** −24.19* (3.626) (7.201) (4.878) (13.59) Log(seed price) −10.30*** −8.829*** −10.44*** −7.098* −14.06*** (2.972) (2.894) (3.084) (4.008) (4.252) Bt single trait 4.265** 9.433*** 17.81*** 16.26*** 24.05** (1.609) (2.870) (5.778) (4.206) (11.31) Stacked variety 11.31*** 17.42*** 26.38*** 29.63*** 35.01*** (2.771) (4.273) (6.604) (5.498) (12.27) Constant 54.83*** 46.84*** 55.62*** 39.97* 74.8*** (16.21) (15.75) (16.84) (21.84) (23.22)

Areas 22 22 22 22 22 Observations 55 55 55 55 55 (Pseudo-)R2 0.296 0.362 0.22 0.13

“Hilly/mountainous” terrain, meanwhile, can limit potential yield and to the estimated conditional logit model with endogenous sorting; the increase farming costs, limiting incentives for farmers to invest in the Supplementary Material provides details. We compute average price extra price premia for more productive varieties. elasticities of demand for the Bt varieties across areas (table S3 in the The seed price coefficient in the base conditional logit model is of Supplementary Material). In the fixed effects conditional logit model the expected sign, though only significant at the 10% level. Seed price is ignoring areawide feedbacks, we estimate that that a 1% price increase omitted from the fixed effects conditional logit model, because as an results in a 1.77% decrease in demand for the Bt varieties (i.e. price- area-level variable it is contained within the estimated area-level fixed elastic demand). When we account for the estimated areawide feedback effects (eq. 4), and thus is included in the second stage regressions (using our preferred IV Tobit model), we estimate that a 1% price in­ (Table 5). The seed price coefficient is always negative and significant crease results in a 0.58% decrease in demand (i.e. price-inelastic de­ in these regressions (in all but one model, at < 1% levels). mand). The second-stage regression estimates in Table 5 exhibit intuitive Turning to the question of whether the above evidence for en­ patterns and appear to confirm our bioeconomic prediction. Across all dogenous sorting is arising due to a bioeconomic feedback from area­ of the specifications in Table 5, we find a statistically significant ne­ wide pest suppression, Table 6 shows results of the FIML model de­ gative feedback effect of areawide Bt deployment on utility from these scribed at the end of section 2.2 and in the Supplementary Material. We varieties in the IV regressions (columns 3 and 8). Moreover, comparing find results highly consistent with our bioeconomic hypothesis. Greater the naïve OLS model (column 1) with the IV models (columns 3–5), we areawide Bt deployment is strongly associated with a significantly re­ see as expected that ignoring endogeneity of the feedback results in a duced perceived likelihood of infestation. Table 6 reports marginal ef­ smaller magnitude (though still negative) feedback effect. (The mag­ fects. For example, using our preferred specification in the last column nitudes of the coefficients are comparable across columns, since the of Table 6, a 10% increase in areawide Bt deployment is estimated to be same fixed effects from the first-stage logit are used as dependent associated with a minimum 10.8% reduction in the likelihood that the variables throughout Table 5.) Our preferred specification, the IV Tobit farmer indicates expecting ACB infestation. This marginal effect is in column 5 that allows for endogeneity and censoring of the fixed ef­ highly robust across specifications controlling for endogeneity and fects, appears to be conservative both in the estimated relative mag­ village-level fixed effects. nitude (discussed below) and statistical precision of the feedback effect, compared to the IV and IVQR models. In terms of IV performance, we 6. Conclusions confirm the Bayer and Timmins instrument appears to have sufficient power, with an F-statistic = 11 yielded by a linear regression of ob­ Bioeconomic feedbacks associated with pest control have important served area-level Bt shares on the IV. Without additional structural implications for agroecological systems. In addition to negative en­ assumptions, statistical tests of exogeneity of the instrument are not vironmental externalities associated with chemical pesticides and the possible, since this IV system is not over-identified. open-access resource issues associated with pesticide resistance, we The estimated coefficients in Table 5 are marginal utilities and draw attention here to the positive externalities associated with area­ preclude direct economic interpretation of magnitudes (although they wide pest suppression spillovers. While previous entomological re­ are comparable across the columns). Two relevant economic quantities search has shown these spillovers to be biologically significant, our can be computed from the estimated coefficients. The first is the ratio econometric analysis is the first to empirically show the potential between the areawide feedback coefficient and the price coefficient, economic significance of these spillovers, in terms of affecting farmers' which can be interpreted as the equivalent variation in utility between demand for pest control technologies. In the presence of pest suppres­ a marginal change in seed price or a marginal change in areawide Bt sion spillovers from Bt crops, farmers who choose not to plant such deployment. In terms of reduced utility to farmers, this calculation crops initially are more likely to continue doing so, as they enjoy the implies a 1% increase in areawide Bt deployment is equivalent to an spillover benefits of neighboring farms' use of Bt crops. Indeed, prior approximately 1.7% increase in seed prices in the IV and IV Tobit media reports in the U.S. have suggested that “farmers are getting models (t-statistics of 1.74 and 1.42 respectively), and in the IVQR savvier about gene shopping” (WSJ 2016), for example avoiding paying model to a 2.82% increase in prices (t-statistic = 1.81). This contrasts the extra technology fee associated with Bt traits in maize targeting the with a much lower equivalent variation estimate of 0.88% estimate Western corn rootworm, due to generally low perceived risks from that implied by the naïve OLS model. pest in the U.S. in recent years. A second quantity of economic interest, as discussed in section 2.1, Beyond Bt crops, our endogenous sorting approach could be are price-elasticities of demand for the Bt varieties. Similarly to our adapted for empirically analyzing other possible bioeconomic feed­ theoretical analysis in eq. (2), we apply the Implicit Function Theorem backs in the control of pests, other bioinvasions and transmissible

8 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883

Table 6 Marginal effects on farmer expecting corn borer infestation (see eq. 7).

Probit IV Probit (FIML)a

(1) (2) (3) (4) (5)

Area-level fraction adopting GM −0.555*** −0.810*** −1.433*** −1.283*** −1.078*** (0.0703) (0.0893) (0.0860) (0.210) (0.194) Own GM adoption (binary) 0.258*** 0.235*** −0.0359 −0.214 (0.0605) (0.0536) (0.218) (0.131) Rolling terrain 0.127*** 0.116** 0.0917* 0.0935* 0.0880* (0.0480) (0.0468) (0.0514) (0.0521) (0.0495) Hilly/mountainous 0.0369 0.0589 0.101* 0.0782 0.0539 (0.0529) (0.0524) (0.0589) (0.0572) (0.0538) Years farming 0.00331** 0.00323** 0.00185 0.00185 0.00171 (0.00150) (0.00148) (0.00142) (0.00180) (0.00171) Cross-eq. correl. 0.540 0.795* (0.372) (0.208) Village fixed effects No No Yes Yes Yes Excluded IVs a All GM premium F-stat 1st stage IVs 18.78*** 17.31*** Observations 515 515 515 515 515 Farms 261 261 261 261 261 Degrees of freedom 4 5 15 31 29 Log-likelihood −329.1 −318.3 −270.9 −509.7 −510.7 Pseudo-R2 0.0704 0.101 0.235

Table 6 Notes: Robust standard errors clustered at grower level in parentheses, * p < 0.10, ** p < 0.05, *** p < 0.01. a FIML IV probit estimated as a correlated bivariate probit system between farmer's expected corn borer infestation and own use of GM maize. Instruments include village-level mean GM seed premium, farm distance to seed source and distance to nearest road. diseases. As reviewed at the outset of this paper, previous applications might infer from our findings that Cilberto et al. underestimate valua­ of the instrumental variables strategy introduced by Bayer and Timmins tions of Bt traits. An important caveat to this inference, however, is the (2007) have ranged over a variety of topics and data structures. Ours is evolution of pest resistance to Bt in the U.S. compared to the Philippines arguably the first paper to show how this approach can be usedto (where Bt resistance was relatively low during the period we analyze). identify explicitly bioeconomic feedbacks. These methods could prove With Bt resistance operating analogously to an open-access resource equally flexible for such applications, which could include study of (Ambec and Desquilbet 2012; Brown 2018), its interaction with the spillovers from vaccination choice (Oster 2016), culling to prevent the ‘local public good’ of areawide pest suppression from Bt is likely to have spread of animal diseases (Gramig and Wolf 2007), or economic re­ complex implications, e.g. for economic valuation and policy. For ex­ sponses to invasive species (Jones 2016). ample, Wan et al. (2012), in finding evidence for areawide pest sup­ The econometric approach and application this paper develops is pression of pink bollworm within Bt cotton systems in China, argue not without limitations. Discussing these limitations and their im­ these pest suppression spillovers may be leveraged to assist Bt re­ plications in detail is instructive for understanding common empirical sistance management. Alternatively, landscape-level pest pressure of issues that arise in studying bioeconomic feedbacks, and for identifying secondary (non-target) pests may increase and encourage further pest future research avenues. First, while we highlight the importance of control expenditures of Bt growers and their neighbors (Kuosmanen landscape-level bioeconomic feedbacks, we lack geolocation data on et al. 2006). In terms of agricultural policy, these interactions suggest farms and fields within the villages surveyed. Future research using that a mix of positive and negative incentives is likely appropriate to such data could model and estimate heterogeneous spillover effects address the multiple externalities generated by these complex ecolo­ across the landscape, characterize network effects, and corroborate gical feedbacks (Lefebvre et al. 2015). results obtained from alternative analytical treatments of pest control Another important limitation of this paper's analysis is the inability spillovers (e.g. using spatial econometrics, Aglasan 2020). to disentangle purely behavioral or sociological feedbacks (e.g. peer Another limitation of our application is the relatively narrow variety effects or social learning) from the alleged pest suppression feedback. of seed products and trait combinations in the data, e.g. compared to There are two concerns here: the validity of our econometric results and U.S. maize farming where the number of seed products available in any the suitability of our endogenous sorting approach in future applica­ given year and market can exceed a dozen (or a hundred, depending on tions to bioeconomic systems. While we cannot rule out potential how a ‘seed product’ is defined, Ciliberto et al. 2019). Applications of econometric bias in our results from unmeasured peer effects and be­ our framework to contexts with greater variety in seed products (and havioral spillovers, we argue these are likely to have a small impact in variation in their availability across geographies) would permit the our application and that any resulting bias is likely to make our con­ study of how bioeconomic feedbacks affect the economic value of a clusions more conservative: Bt crops were available for a number of wide variety of crop traits. Indeed, Ciliberto et al. (2019) use the same years and widely deployed in the Philippines prior to the study period, underlying discrete-choice framework used in this paper – with a more and information about these crops appears to have been disseminated flexible and expansive model of farmer choice but without bioeconomic widely. Likewise, the stacked trait variety was deployed between the spillovers or feedbacks – to obtain economic valuations of GM traits in first and second farmer surveys, and (as shown above) achieved rapid maize and soy in the U.S. adoption by the time of the second survey. Additionally, to the extent An open research question is how accounting for the type of bioe­ that such behavioral spillovers are present and positive (with farmers conomic spillovers we identify here, reestimated for the U.S., would being more likely to mimic their neighbors in deploying Bt), they likely affect Ciliberto et al.'s valuations. At a basic level, because our analysis bias our negative estimated pest suppression feedback towards zero. empirically shows that farmers to some extent ‘free-ride’ on others' This means that the magnitude we estimate on this feedback is likely adoption of the Bt trait, and because Ciliberto et al.'s valuations are conservative. based on estimated farmers' willingness to pay for different traits, one In terms of the suitability of an endogenous sorting approach for

9 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883 future bioeconomic analyses, separating out the behavioral feedbacks Appendix A. Supplementary data requires a dedicated identification strategy on top of the endogenous sorting models we study here. A number of recent papers and works in Supplementary data to this article can be found online at https:// progress introduce novel strategies using field experiments for identi­ doi.org/10.1016/j.ecolecon.2020.106883. fying peer and network effects in technology adoption Magnan( et al. 2015; Beaman et al. 2018), some using the same discrete choice, References random utility framework used here (Dickinson et al. 2018; Guiteras et al. 2019). Combining such behavioral experiments (or quasi-experi­ Morallo-Rejesus, Belen G., Punzalan, E., 2002. Mass Rearing and Field Augmentation of ments) with study of the underlying ecology can create more empiri­ the Earwig, Euborellia Annulata, against Asian Corn Borer. Afidchao, M.M., Musters, C., de Snoo, G.R., 2013. Asian corn borer (ACB) and non-ACB cally grounded models of bioeconomic systems (Brown 2018). pests in GM corn (Zea mays L.) in the Philippines. Pest Manag. Sci. 69 (7), 792–801. Our analysis could also benefit from the addition of instrumental Available at: https://doi.org/10.1002/ps.3471 [Accessed September 11, 2017]. measurements on pest density or physical crop damage, to corroborate Aglasan, S., 2020. A Spatial Econometric Analysis of Unexpected Rootworm Damage in Bt Corn Fields. In: B. K. Goodwin (Dissertation Chair), ed. Three Essays on the the areawide pest suppression implied by our analysis of farmer seed Econometric Analysis of Biotic and Abiotic Risk in the Cultivation of Genetically choices. This paper addresses the lack of such measurements by citing Modified Corn. North Carolina State University, Raleigh, NC, pp. 126. Available at: extensive areawide suppression of the European corn borer (a close https://repository.lib.ncsu.edu/handle/1840.20/38151. relative of ACB) with Bt maize that has been observed in U.S. Ambec, S., Desquilbet, M., 2012. Regulation of a spatial externality: refuges versus tax for managing Pest resistance. Environ. Resour. Econ. 51 (1), 79–104. Available at: (Hutchison et al., 2010), and by showing that Filipino maize farmers' https://doi.org/10.1007/s10640-011-9489-3 [Accessed September 22, 2015]. perceived ACB infestation risk appears to significantly decrease in re­ Bayer, P., Timmins, C., 2005. On the equilibrium properties of locational sorting models. sponse to greater village-level use of Bt. Nevertheless, having direct pest J. Urban Econ. 57 (3), 462–477. Bayer, P., Timmins, C., 2007. Estimating equilibrium models of sorting across locations. density or damage measurements would undoubtedly improve our ap­ Econ. J. 117 (518), 353–374. proach. Whereas the reader could view this paper as a ‘partial in­ Beaman, L.A., BenYishay, A., Magruder, J., Mobarak, A.M., 2018. “Can Network Theory- formation’ analysis of endogeneity in farmer demand for Bt maize, the Based Targeting Increase Technology Adoption?” NBER Working Paper Series No. 24912. addition of pest density observations could be used for a ‘full in­ Bernasco, W., de Graaff, T., Rouwendal, J., Steenbeek, W., 2017. Social interactions and formation,’ simultaneous estimation of farmer seed choices and ex­ crime revisited: an investigation using individual offender data in Dutch neighbor­ plicitly modeled pest population dynamics. Estimating such a model hoods. Rev. Econ. Stat. 99 (4), 622–636. Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices in market equilibrium. would facilitate empirically based counterfactual or predictive bioeco­ Econometrica 63 (4), 841–890. nomic simulations, e.g. of how pest populations react to farmer re­ Birol, E., Villalba, E.R., Smale, M., 2008. Farmer preferences for milpa diversity and sponses to changes in pricing of Bt traits or in the bundling of Bt with genetically modified maize in Mexico: a latent class approach. Environ. Dev. Econ. 14 (04), 521. Available at: http://journals.cambridge.org/abstract_ other seed traits such as herbicide tolerance. The addition of explicit S1355770X08004944, Accessed date: 15 March 2016. pest density data could also be useful for separating out bioeconomic Brown, Z.S., 2018. Voluntary programs to encourage compliance with refuge regulations and previously discussed social learning spillovers. for pesticide resistance management: evidence from a quasi-experiment. Am. J. Agric. Table 5 notes: Standard errors in parentheses, * p < 0.10, ** Econ. 100 (3), 844–867. Campbell, J.R., Hopenhayn, H.A., 2005. Market Size Matters. J. Ind. Econ. 53 (1), 1–25. p < 0.05, *** p < 0.01. Unless otherwise noted, we report jackknife Available at: https://doi.org/10.1111/j.0022-1821.2005.00243.x [Accessed January standard errors clustered at the area level and obtained from joint es­ 15, 2018]. timation of first-stage (fixed effects conditional logit, Table 4) and Carrière, Y., Brown, Z.S., Downes, S.J., Gujar, G., Epstein, G., Omoto, C., Storer, N.P., a Mota-Sanchez, D., Søgaard Jørgensen, P., Carroll, S.P., 2020. Governing evolution: a second-stage regressions. Area Bt fraction instrumented using Bayer socioecological comparison of resistance management for insecticidal transgenic Bt and Timmins instrument. b IVQR is an instrumental variable quantile crops among four countries. Ambio 49 (1), 1–16. regression, using the quantile treatment effects model ofChernozhukov Chambers, R.G., Karagiannis, G., Tzouvelekas, V., 2010. Another look at pesticide pro­ ductivity and Pest damage. Am. J. Agric. Econ. 92 (5), 1401–1419. Available at: and Hansen (2005, 2006). Due to invalidity of jackknifed standard er­ https://doi.org/10.1093/ajae/aaq066. rors for quantile regression (Shao and Wu 1989) and complex data Chernozhukov, V., Hansen, C., 2005. An IV model of quantile treatment effects. structure for bootstrap resampling jointly over the discrete choice first Econometrica 73 (1), 245–261. Available at: https://doi.org/10.1111/j.1468-0262. 2005.00570.x [Accessed March 15, 2016]. stage and area-level second stage, IVQR standard errors are obtained Chernozhukov, V., Hansen, C., 2006. Instrumental quantile regression inference for from bootstrap resampling only over the second-stage data (ignoring structural and treatment effect models. J. Econ. 132 (2), 491–525. Available at: first-stage measurement error in the δ’s). http://www.sciencedirect.com/science/article/pii/S0304407605000643, Accessed date: 3 January 2018. Ciliberto, F., Moschini, G.C., Perry, E.D., 2019. Valuing product innovation: genetically engineered varieties in US corn and soybeans. RAND J. Econ. 50 (3), 615–644. Declaration of Competing Interest Clark, C.W., 2010. Mathematical Bioeconomics: The Mathematics of Conservation 3rd Ed. John Wiley & Sons. Connor, L.Q., 2017. Post-Adoption Impacts of GM Corn on Farm Decisions in the The authors declare that they have no known competing financial Philippines. Available at: https://repository.lib.ncsu.edu/handle/1840.20/34356, interests or personal relationships that could have appeared to influ­ Accessed date: 23 January 2018. Culliney, T.W., 2014. “Crop Losses to Arthropods.” In Integrated Pest Management. ence the work reported in this paper. Dordrecht: Springer Netherlands. pp. 201–225. Available at: https://doi.org/10. 1007/978-94-007-7796-5_8 [Accessed October 21, 2016]. Deutsch, C.A., Tewksbury, J.J., Tigchelaar, M., Battisti, D.S., Merrill, S.C., Huey, R.B., Acknowledgements Naylor, R.L., 2018. Increase in crop losses to insect pests in a warming climate Downloaded from. Available at: http://science.sciencemag.org/, Accessed date: 19 August 2019. Funding for data collection was provided by the International Food Dickinson, K.L., Dalaba, M., Brown, Z.S., Alirigia, R., Coffey, E.R., Mesenbring, E., Policy Research Institute (IFPRI). This work was also supported by the Achazanaga, M., Agao, D., Ali, M., Kanyomse, E., Awaregya, J., Adagenera, C.A., Aburiya, J.B.A., Gubilla, B., Oduro, A.R., Hannigan, M.P., 2018. Prices, peers, and USDA National Institute of Food and Agriculture HATCH project perceptions (P3): study protocol for improved biomass cookstove project in northern #NC02520, the NCSU Genetic Engineering & Society Center, and the Ghana. BMC Public Health 18 (1), 1209. Available at: https://doi.org/10.1186/ NCSU Center for Environmental and Natural Resource Economics and s12889-018-6116-z [Accessed July 15, 2020]. Dively, G.P., Venugopal, P.D., Bean, D., Whalen, J., Holmstrom, K., Kuhar, T.P., Doughty, Policy. The authors have no conflicts of interest to declare in relation to H.B., Patton, T., Cissel, W., Hutchison, W.D., 2018. Regional pest suppression asso­ this research. We are grateful for the useful feedback received from ciated with widespread Bt maize adoption benefits vegetable growers. In:Proceedings seminar participants at the University of Wisconsin, Madison; the of the National Academy of Sciences of the United States of America 115(13):3320–3325. University of Nebraska, Lincoln; participants in the BIOECON 2017 Available at, . http://www.ncbi.nlm.nih.gov/pubmed/29531049, Accessed date: 9 April 2018. meetings in Tilburg; the USDA Economic Research Service; Chris Epanchin-Niell, R.S., Wilen, J.E., 2015. Individual and cooperative management of in­ Timmins; as well as two anonymous reviewers. vasive species in human-mediated landscapes. Am. J. Agric. Econ. 97 (1), 180–198.

10 Z.S. Brown, et al. Ecological Economics 180 (2021) 106883

Available at: https://academic.oup.com/ajae/article/97/1/180/2737502, Accessed March 11, 2016]. date: 9 September 2020. Murdock, J., 2006. Handling unobserved site characteristics in random utility models of FELKL, G., 1988a. Effect of detasseling corn plants on egg deposition, larval infestation recreation demand. J. Environ. Econ. Manag. 51 (1), 1–25. Available at: http:// and tunneling damage of Asian corn borer, Ostrinia furnacalis Guenée (Lepidoptera: www.sciencedirect.com/science/article/pii/S0095069605000574, Accessed date: 31 Pyr- alidae), in the Philippines. J. Plant Dis. Prot. 95, 476–485. October 2017. FELKL, G., 1988b. Economic aspects of detasseling corn plants and insecticide use to Nafus, D.M., Schreiner, I.H., 1991. Review of the biology and control of the Asian corn control Asian corn borer, Ostrinia furnacalis Guenée (Lepidoptera: Pyralidae), in the borer, Ostrinia furnacalis (Lep: Pyralidae). Tropical Pest Management 37 (1), 41–56. Philippines. J. Appl. Ecol. 105, 37–386. Available at: https://doi.org/10.1080/09670879109371535 [Accessed November 2, Fernandez-Cornejo, J., Wechsler, S., Livingston, M., Mitchell, L., 2014. Genetically en­ 2017]. gineered crops in the United States. No. ERR-162 Available at: www.ers.usda.gov, Oerke, E.-C., 2006. Crop losses to pests. J. Agric. Sci. 144 (01), 31. Available at: http:// Accessed date: 19 August 2019. journals.cambridge.org/abstract_S0021859605005708, Accessed date: 19 October Gould, F., Brown, Z.S., Kuzma, J., 2018. Wicked evolution: can we address the socio­ 2015. biological dilemma of pesticide resistance? Science (New York, N.Y.) 360 (6390), Oster, E., 2016. Does Disease Cause Vaccination? Disease Outbreaks and Vaccination 728–732. Available at: http://www.ncbi.nlm.nih.gov/pubmed/29773742, Accessed Response. date: 8 June 2018. PestNet, 2017. PestNet. In: Pacific Pests and Pathogens - Fact Sheets: Maize Asian corn Gouse, M., Sengupta, D., Zambrano, P., Zepeda, J.F., 2016. Genetically modified maize: borer, Available at: https://www.pestnet.org/fact_sheets/maize_asian_corn_borer_ less drudgery for her, more maize for him? Evidence from smallholder maize farmers 115.htm, Accessed date: 16 July 2020. in South Africa. World Dev. 83, 27–38. Available at: https://www.sciencedirect. Rodriguez, T., 2014. The revised refuge requirement for biotech corn farms – agriculture com/science/article/pii/S0305750X16000498, Accessed date: 23 January 2018. monthly. In: Agriculture Monthly, Available at: https://www.agriculture.com.ph/ Gramig, B.M., Wolf, C.A., 2007. Estimating within-herd preventive spillovers in livestock 2018/01/31/the-revised-refuge-requirement-for-biotech-corn-farms/, Accessed date: disease management. Am. J. Agric. Econ. 89 (5), 1219–1225. Available at: https:// 16 July 2020. doi.org/10.1111/j.1467-8276.2007.01087.x [Accessed December 31, 2019]. Sanglestsawai, S., Rejesus, R.M., Yorobe, J.M., 2014. Do lower yielding farmers benefit Groeneveld, R.A., Wesseler, J., Berentsen, P.B.M., 2013. Dominos in the dairy: an analysis from Bt corn? Evidence from instrumental variable quantile regressions. Food Policy of transgenic maize in Dutch dairy farming. Ecol. Econ. 86, 107–116. 44, 285–296. Available at: http://linkinghub.elsevier.com/retrieve/pii/ Guiteras, R., Levinsohn, J., Mobarak, A.M., 2019. “Demand Estimation with Strategic S0306919213001395, Accessed date: 10 March 2014. Complementarities: Sanitation in Bangladesh.” CEPR Discussion Paper No. DP13498. Shao, J., Wu, C.F.J., 1989. A General Theory for Jackknife Variance Estimation. Hicks, R.L., Horrace, W.C., Schnier, K.E., 2012. Strategic substitutes or complements? The Available at: https://www.jstor.org/stable/pdf/2241717.pdf?refreqid=excelsior game of where to fish. J. Econ. 168 (1), 70–80. Available at: http://www. %3Ac3f5f0a02941dfd71a6b53208a111495, Accessed date: 26 September 2018. sciencedirect.com/science/article/pii/S0304407611001709, Accessed date: 8 April Shirai, Y., 1998. Laboratory evaluation of flight ability of the oriental corn borer, Ostrinia 2015. furnacalis (Lepidoptera: Pyralidae). Bull. Entomol. Res. 88, 327–333. Available at: Hutchison, W.D., Burkness, E.C., Mitchell, P.D., Moon, R.D., Leslie, T.W., Fleischer, S.J., http://cabweb.org/PDF/BER/BER88-3/327.pdf, Accessed date: 2 November 2017. Abrahamson, M., Hamilton, K.L., Steffey, K.L., Gray, M.E., Hellmich, R.L., Kaster, Skevas, T., Fevereiro, P., Wesseler, J., 2010. Coexistence regulations and agriculture L.V., Hunt, T.E., Wright, R.J., Pecinovsky, K., Rabaey, T.L., Flood, B.R., Raun, E.S., production: a case study of five Bt maize producers in Portugal. Ecol. Econ. 69 (12), 2010. Areawide suppression of European corn borer with Bt maize reaps savings to 2402–2408. non-Bt maize growers. Science 330 (6001), 222–225. Available at. http://www.ncbi. Smart, R.D., Blum, M., Wesseler, J., 2017. Trends in approval times for genetically en­ nlm.nih.gov/pubmed/20929774, Accessed date: 12 August 2014. gineered crops in the United States and the European Union. J. Agric. Econ. 68 (1), Jones, B.A., 2016. Work more and play less? Time use impacts of changing ecosystem 182–198. Available at: https://doi.org/10.1111/1477-9552.12171 [Accessed services: the case of the invasive emerald ash borer. Ecol. Econ. 124, 49–58. October 5, 2020]. Kuosmanen, T., Pemsl, D., Wesseler, J., 2006. Specification and estimation of production Timmins, C., 2007. If you cannot take the heat, get out of the Cerrado? Recovering the functions involving damage control inputs: a two-stage, semiparametric approach. equilibrium amenity cost of nonmarginal climate change in Brazil. J. Reg. Sci. 47 (1), Am. J. Agric. Econ. 88 (2), 499–511. Available at: https://doi.org/10.1111/j.1467- 1–25. Available at: https://doi.org/10.1111/j.1467-9787.2007.00497.x [Accessed 8276.2006.00874.x [Accessed July 15, 2020]. April 8, 2015]. Gerpacio, R.V., Labios, J.D., Labios, R.V., Diangkinay, E.I., 2004. Maize in the Philippines: Timmins, C., Murdock, J., 2007. A revealed preference approach to the measurement of Production Systems. Constraints, and Research Priorities. CIMMYT. congestion in travel cost models. J. Environ. Econ. Manag. 53 (2), 230–249. Available Lefebvre, M., Langrell, S.R.H., Gomez-y-Paloma, S., 2015. Incentives and policies for at: http://www.sciencedirect.com/science/article/pii/S0095069606000817, integrated pest management in Europe: a review. Agron. Sustain. Dev. 35 (1), 27–45. Accessed date: 21 April 2015. Available at: https://doi.org/10.1007/s13593-014-0237-2 [Accessed April 26, Train, K.E., 2009. Discrete Choice Methods with Simulation. Cambridge University Press. 2017]. Useche, P., Barham, B.L., Foltz, J.D., 2009. Integrating technology traits and producer Magnan, N., Spielman, D.J., Lybbert, T.J., Gulati, K., 2015. Leveling with friends: social heterogeneity: a mixed-multinomial model of genetically modified corn adoption. networks and Indian farmers’ demand for a technology with heterogeneous benefits. Am. J. Agric. Econ. 91 (2), 444–461. Available at: https://doi.org/10.1111/j.1467- J. Dev. Econ. 116, 223–251. 8276.2008.01236.x [Accessed November 19, 2014]. Mbah, M.L.N., Forster, G.A., Wesseler, J.H., Gilligan, C.A., 2010. Economically optimal Wan, P., Huang, Y., Tabashnik, B.E., Huang, M., Wu, K., 2012. “The halo effect: sup­ timing for crop disease control under uncertainty: an options approach. J. R. Soc. pression of pink bollworm on non-Bt cotton by Bt cotton in China” A. Bravo, ed. PLoS Interface 7 (51), 1421–1428. Available at: https://doi.org/10.1098/rsif.2010.0056 ONE 7 (7). https://doi.org/10.1371/journal.pone.0042004. e42004. Available at:. [Accessed July 15, 2020]. [Accessed August 21, 2019]. McFadden, D., 1978. Modeling the choice of residential location. Transp. Res. Rec. 673. Werle, R., Oliveira, M.C., Jhala, A.J., Proctor, C.A., Rees, J., Klein, R., 2018. Survey of Melitz, M.J., Ottaviano, G.I.P., 2008. Market size, trade, and productivity. Review of Nebraska Farmers’ Adoption of Dicamba-Resistant Soybean Technology and Dicamba Economic Studies 75 (1), 295–316. Available at: https://doi.org/10.1111/j.1467- off-Target Movement. Cambridge University Press Available at: https://www. 937X.2007.00463.x [Accessed January 15, 2018]. cambridge.org/core/product/identifier/S0890037X18000623/type/journal_article, Mendoza, M.S., Rosegrant, M.W., 1995. “Pricing Behavior in Philippine Corn Markets: Accessed date: 15 August 2019. Implications for Market Efficiency.” Research Report- International Food Policy WSJ, 2016. Behind the Monsanto Deal, Doubts About the GMO Revolution. The Wall Research Institute:11–79. Street Journal Available at: http://www.wsj.com/articles/behind-the-monsanto- Milne, A.E., Bell, J.R., Hutchison, W.D., van den Bosch, F., Mitchell, P.D., Crowder, D., deal-doubts-about-the-gmo-revolution-1473880429. Parnell, S., Whitmore, A.P., 2015. The effect of Farmers’ decisions on pest control Yorobe, J.J.M., Quicoy, C.B., 2006. Economic impact of Bt corn in the Philippines. with Bt crops: a billion Dollar game of strategy. PLoS Comput. Biol. 11 (12), Philippine Agricultural Scientist 89 (258–267). e1004483 Available at: https://doi.org/10.1371/journal.pcbi.1004483 [Accessed

11